Financial Industry Agent 100 Skills: From General AI to Digital Employees – A Practical Guide
By 2026, AI adoption in China’s financial sector has crystallized into concrete numbers: 7‑day iteration cycles, over 2,500 deployed agents, and end‑to‑end process re‑engineering. Large models plus agentic systems have proven the arrival of what some call the “Loopy Era” – agents that can autonomously design experiments, refine models, and iterate without human supervision. This isn’t a roadmap anymore; it’s what’s happening on the ground.
Yet one question keeps coming up in every bank, insurer, and asset manager I talk to: How do you make a general‑purpose AI agent actually useful – and trustworthy – in a heavily regulated, high‑stakes financial environment?
The answer isn’t about picking a bigger model. It’s about engineering a different kind of system – one that encodes decades of domain expertise into reusable, auditable digital assets. This post breaks down the architecture, the practical skills, and the real‑world deployment journey based on a comprehensive industry playbook that covers 100+ production‑ready financial skills across ten core roles.
Why the Agent Era Is Already Here, Not Coming
The simplest measure of a tech shift is whether people pay real money for it. That test is passed.
OpenAI’s annual recurring revenue (ARR) went from roughly 20B by the end of 2025 – a 10x jump in two years. Anthropic grew even faster, from 30B by early 2026. These aren’t startup valuations; they’re revenue. Enterprises are signing contracts, renewing them, and expanding usage.
Then look at product velocity. Throughout 2025 and into 2026, OpenAI shipped a major model or product update every 6–8 weeks – GPT‑4.1, o3 series, GPT‑5, GPT‑5.5, plus agent infrastructure like Operator and the Assistants API. Anthropic matched that pace with Claude 3, 3.5, 4, 4.5, and Sonnet 4.6, each bringing meaningful capability jumps in coding, long‑context reasoning, and tool use.
The hardest signal comes from capital expenditure. In 2026, Google, Microsoft, Amazon, and Meta together are projected to spend around **190 billion. When the world’s most sophisticated business leaders collectively bet that kind of money, they are not gambling – they are placing a structural bet on a paradigm shift.
For financial institutions, the implication is stark: waiting is becoming exponentially more expensive. Every quarter of hesitation means your competitors have accumulated another round of data, feedback, and refined agent capabilities. History shows that in every major computing transition, the “can’t see it, won’t believe it, don’t understand it, too late” cycle doesn’t kill the smallest players – it kills the ones who were too proud to move early.
The Three‑Engine Positive Feedback Loop
Why is this AI wave growing exponentially rather than linearly? Because three engines – model capability, agent products, and ecosystem – are feeding off each other.
Engine 1: The Model Foundation
By early 2026, reasoning models reached full maturity. GPT‑5.5 scored 82.6% on SWE‑bench Verified, surpassing the average human expert for the first time on software engineering tasks. Claude Opus 4.7 supports a 1‑million‑token context window and is designed to “work autonomously for hours on the hardest coding and research tasks.” On the domestic side, Qwen3.6‑Plus scored 61.6% on Terminal‑Bench 2.0, beating several previous‑generation overseas flagship models.
What really matters is the emergence of “slow thinking” – the model’s ability to break down complex problems, verify intermediate steps, and correct itself. This turns AI from a chatty assistant into a credible partner for financial analysis, legal reasoning, and scientific research.
Engine 2: Agent Products
Claude Code’s evolution is a perfect case study. In early 2025, it was a code‑suggestion tool. By Q1 2026, it had become a fully autonomous agent: Loop Mode runs up to 50 concurrent scheduled tasks; Computer Use manipulates macOS and browsers directly; Channels let multiple agent instances collaborate. It now handles the entire development lifecycle – from requirement understanding to deployment – with the productivity of a senior engineer working full‑time.
Engine 3: Ecosystem
The Model Context Protocol (MCP) reached nearly 97 million monthly downloads by March 2026, with over 10,000 registered MCP servers. The Agent‑to‑Agent (A2A) protocol, released by Google, solved the “how agents talk to each other” problem. Together, they give agents a standardized way to connect to tools, data sources, and other agents.
The key insight for financial firms: the earlier you join this loop, the more proprietary interaction data you accumulate. That data becomes a barrier that later entrants cannot easily replicate. In a compounding system, time is your most irreplaceable asset.
From “Financial‑Grade Cloud‑Native” to “Financial‑Grade AI‑Native”
Adding a model API to your existing systems won’t get you there. Just like moving a monolithic app to a VM didn’t deliver elasticity – you needed microservices, containers, and DevOps. The same logic applies here. AI‑native isn’t about adding AI; it’s about rethinking how business systems handle interaction, decision‑making, and knowledge.
Four mechanisms work together to make this real:
LUI (Language User Interface) – “Carbon sees the screen, silicon reads the intent”
Under the old GUI paradigm, a relationship manager preparing for a client meeting would log into the CRM, then the asset management system, then the market data terminal, then the document system – four systems, dozens of clicks, 30+ minutes.
With LUI, they simply say: “Prepare tomorrow’s meeting materials for Mr. Zhang.” The agent automatically queries the CRM for his profile, pulls portfolio data, overlays market conditions, generates a personalized proposal, and drafts the talking points. Total time: about two minutes.
China Construction Bank’s 2026 mobile banking update is a real‑world example – they describe it as “not a voice replacement for clicks, but a service‑paradigm reconstruction based on large‑model understanding.”
Hooks – Guardrails for Compliance
Compliance isn’t a afterthought; it’s built into the execution path. Hooks are preset interception points. When an agent’s workflow hits a threshold – transaction size, sensitive data access, regulatory red line – the hook pauses execution, shows the full reasoning chain (data sources, rules applied, intermediate judgments), and waits for human approval.
Credit approval scenario: the agent completes the initial screening and risk scoring, then triggers a hook at the final decision step. The credit officer sees every piece of evidence the agent used and makes the final call.
The design philosophy: give the agent full autonomy inside the “safe zone,” but require human oversight at every “sensitive zone.”
Skill – Encoding Domain Expertise
The most valuable asset in any financial institution isn’t just data – it’s the tacit knowledge in the heads of seasoned professionals. That judgment that can’t be written into a rulebook, the instinct that flags a seemingly normal transaction as suspicious.
The Skill mechanism lets business experts describe their workflows – triggers, steps, quality standards, edge cases – in natural language. The system then turns that description into a structured, executable, and auditable skill that any agent can use.
Example – credit limit uplift for specialized SMEs: a senior banker’s rule of thumb – “check industry whitelist, assess core financial ratios, add innovation bonus, calculate uplift” – becomes a skill. Now every agent applies the same logic, consistently, every time. When that banker retires, the expertise doesn’t leave.
CLI – Bridging the Legacy Systems
Large banks run hundreds of legacy systems – core banking, CRM, data warehouses, risk, compliance, asset management. They speak different protocols. Instead of rewriting everything, the AI‑native approach “CLI‑fies” each system: provide a standard command interface (or wrap it with MCP) so the agent can query and invoke across systems using a unified language.
The strategic payoff: you don’t have to wait for a “perfect unified platform.” You can modernize incrementally, system by system, while the agent already starts delivering value.
Why a General Agent + N Skills Will Replace N Specialized Agents
There are two architectural paths: build a separate agent for every business scenario, or build one general agent with pluggable skills. The data and logic overwhelmingly favor the second.
Cost
Assume one specialized agent costs 10 person‑months to build, and one skill costs 0.5 person‑months. For 20 scenarios:
-
Specialized: 200 person‑months -
General + skills: 30 (base) + 20×0.5 = 40 person‑months
That’s an 80% cost saving. With 100 scenarios, the gap widens to more than 10x.
Cross‑domain Collaboration
Real financial tasks are rarely siloed. A relationship manager preparing for a CFO meeting needs: company profiling (corporate banking), relationship history (CRM), industry dynamics (research), product solutions (product), and compliance checks (compliance).
Under specialized agents, five different agents must coordinate – passing data, syncing state, and reconciling outputs. Under a general agent, it simply calls five skills in sequence, maintaining a single, coherent context.
Research from Google and MIT (arXiv:2512.08296) found that multi‑agent systems degraded performance by an average of 3.5% overall. For tasks requiring sequential reasoning and context continuity – exactly what financial work demands – the drop was 39%–70%.
Evolution Speed
When regulations change, a specialized agent requires re‑architecting, re‑developing, re‑testing, and re‑deploying – weeks to months. A general agent just needs one skill updated and registered – hours to days.
Specialized agents are fixed‑capability sets; a general agent is a capability‑learning platform. That’s the fundamental difference.
Ecosystem Effect
Think of the smartphone transition. Feature phones had dedicated apps for each function. The iPhone gave us a general platform (iOS) and an open app store – turning capability creation from a centralized effort into a decentralized, exponential ecosystem.
The general agent is the iPhone of this era; skills are the apps. The organization that builds the richest skill ecosystem wins, not the one with the biggest base model.
The 10 Essential Elements of a Financial‑Grade General Agent
What does a production‑ready financial agent actually need? Here are the ten non‑negotiable elements.
1. Resilient Execution – Goal‑Directed, Self‑Planning
Traditional RPA and BPM follow predefined flowcharts. They work for stable processes but break when the situation requires judgment. A resilient agent doesn’t follow a fixed script – it receives a goal, then iterates through Think → Plan → Act → Observe, adjusting its path as conditions change.
For a credit assessment, the agent breaks down the goal into: gather company info → extract financials → benchmark against industry → identify risk factors → compute rating → generate report. If a data source is unavailable, it finds an alternative. If it hits a contradiction, it backtracks.
The model must support: deep reasoning, long‑range planning (1M token context), multimodal understanding (PDFs, scanned docs, charts), precise tool use (98%+ accuracy), and self‑correction.
2. Skill Encapsulation – Domain Knowledge Digitized
A skill is not an API wrapper. It’s a standardized contract that includes: metadata (name, version), trigger conditions, step‑by‑step instructions, tool declarations, output schema, and built‑in compliance rules (e.g., “never give a positive rating to a shell company”).
Five skill patterns emerge: orchestration (multi‑step workflows), retrieval (document Q&A), analysis (multi‑dimensional scoring), generation (report writing), and monitoring (real‑time anomaly detection).
3. Deep Companion – Persistent, Personalized Understanding
Most financial services are “amnesic” – every interaction starts from scratch. A general agent builds a three‑layer memory:
-
Working memory – current session context -
Episodic memory – specific events over weeks/months -
Semantic memory – long‑term user profile (risk tolerance, investment preferences, life stage)
A user digital twin integrates cognitive, emotional, behavioral, and life‑stage dimensions. It doesn’t just record what the client did – it infers why, and predicts what they might need next.
The highest form is proactive advising – the agent monitors market signals, behavioral changes, and life events, then surfaces pre‑vetted recommendations before the client even asks.
4. Interactive Experience – Natural, Frictionless
The UI paradigm is shifting from GUI (graphical) to LUI (language) . Users don’t have to navigate menus; they just say what they want. “Find a fund similar to the one I bought last year but with lower fees” – that’s a single sentence, not a series of drop‑downs.
Generative UI takes it further: the interface is rendered on‑the‑fly based on the user’s expertise and context. A seasoned investor sees a detailed dashboard with Greeks and factor exposures; a retail client sees a simple gain/loss summary. The A2UI protocol provides the standard for this.
5. Controlled Evolution – Getting Smarter, Safely
The agent must evolve, but not wildly. Three layers:
-
Agent‑level (real‑time): tweak personality files (IDENTITY.md, MEMORY.md) based on feedback – e.g., if a user often asks for simpler language, the agent adapts. -
Skill‑level (daily): the agent reviews its own execution traces, extracts reusable patterns, and proposes new skills – but they go through an isolation sandbox, automated and human validation, and gradual rollout before going live. -
Model‑level (weekly/monthly): fine‑tuning or distillation using production feedback and synthetic data – with extended grayscale monitoring and full rollback capability.
6. Legacy Enablement – Bridge, Don’t Replace
Through CLI and MCP bridges, existing systems become callable without code changes. A CLI bridge can turn a shell script into a tool with just a descriptor file. MCP provides protocol‑grade integration for APIs and microservices. This lets the agent start adding value immediately, without waiting for a multi‑year modernization program.
7. Compliance First – Embedded, Not Bolted‑On
Three lines of defense:
-
During reasoning: persona redlines (hard‑coded prohibitions), SOP gates (mandatory step sequences), hooks (real‑time rule checks), and Human‑on‑the‑Loop (HOTL) for high‑risk actions. -
At input/output: bidirectional compliance validation, fact‑anchoring (every claim traceable to a source), suitability checks (product matches client risk profile), and templated output constraints. -
After the fact: full decision‑chain logging and explainability – every output comes with the reasoning and sources that led to it.
8. Technical Security – Plan, Approve, Execute
Security is a three‑stage pipeline:
-
Plan: detect sensitive data, filter malicious inputs (prompt injection), enforce intent whitelists. -
Approve: inherit user permissions, enforce tool whitelists, require hook approval for sensitive operations. -
Execute: run in sandboxed containers, validate model outputs against schemas, use CLI controls for high‑privilege actions, and isolate credentials.
9. Robust Reliability – Service Continuity and Intelligent Fault Tolerance
Reliability is not just uptime; it’s about intelligent resilience. The agent must handle step‑level failures with retries, fallbacks, and checkpoint recovery. Long‑running tasks (hours or days) must be resumable from the last checkpoint. For model‑level errors, a multi‑model routing layer provides automatic switching. If the agent cannot produce a reliable answer, it must explicitly say so, rather than guess.
10. Efficiency Optimization – Mixed‑Model Architecture for ROI
Not every request needs the biggest model. A smart router classifies tasks:
-
~70% of simple queries (FAQ, balance checks) go to small models or rules. -
~25% of moderate tasks (product comparison) go to medium domain models. -
~5% of deep reasoning (investment proposals) go to large models.
Combined with context pruning, prompt caching, KV cache reuse, and output length control, this cuts inference costs by 60%–70% without sacrificing critical quality. Deployment can be hybrid – public cloud MaaS for non‑sensitive work, dedicated cloud instances for core business, and fully on‑prem for the most sensitive data.
Building an AI Bank Wealth Manager: A Step‑by‑Step Walkthrough
Let’s make this concrete. Here’s how you’d create an AI bank wealth manager on a production platform (codename “FinX”).
Step 1: Identity Definition
-
Name: AI Bank Wealth Manager -
Description: Specializes in personalized asset allocation based on risk tolerance, life stage, and financial goals – across funds, insurance, and trusts. -
Soul.md: Defines tone and values – “You are a financial partner rooted in the client’s life, not a salesperson, not a product pusher. You believe in ‘slow allocation, fast response’ – no chasing hot products, no promises of returns, only data‑backed recommendations.”
Step 2: Skill Assembly
Load skills from a marketplace or upload a .zip package. The wealth manager’s skill set covers the entire client lifecycle:
-
Client KYC profiling -
Opportunity scanning -
Market event interpretation -
Portfolio health check -
Asset allocation optimization -
Comprehensive financial planning (simulating a CFP® workflow)
Step 3: Capabilities and Memory
Enable scheduling, code sandboxing, etc. Memory automatically captures key client attributes (risk tolerance, asset size, preferences) during conversation, maintains them persistently, and updates when new information conflicts.
Step 4: Data Sources
Built‑in connectors for major financial data providers (Hengsheng, East Money, Yingmi) and support for custom MCP/HTTP endpoints to internal CRM, portfolio systems, and product databases.
Step 5: Permissions and Guardrails
Three rule types: ALLOW, ASK (human confirmation), DENY. For example, you can set a rule that “risk‑level matching requires human confirmation” – the agent will pause and ask for approval before recommending a product that might be unsuitable.
Self‑Evolution in Action
Level 1 example: The agent sends a daily market brief to a client, who complains it’s too generic. The agent asks follow‑up questions to learn the client holds a specific ETF, then provides a targeted analysis. The client is satisfied. The agent updates its MEMORY.md with: “This client prefers analysis tied to their specific holdings.”
Level 2 example: A user asks about a mutual fund. The agent uses the fund‑deep‑research skill and gives a detailed, visual report. The user then says, “It would be helpful to include related product recommendations.” The agent identifies that this feedback should modify the skill itself, updates the SKILL.md file accordingly, and from then on every fund analysis includes related product suggestions.
Level 3 example: The platform’s model‑flywheel system collects production feedback daily, runs periodic supervised fine‑tuning and reinforcement learning, and redeploys improved domain models after offline evaluation and extended grayscale testing.
The 100+ Financial Skill Landscape Across 10 Roles
The full playbook covers ten core roles with a total of over 100 skills. Here’s a high‑level view:
-
AI Wealth Manager (Retail) – 28 skills: KYC, opportunity scan, market interpretation, portfolio diagnosis, asset allocation, financial planning, client care scripts, etc. -
AI Corporate Relationship Manager – 8 skills: visit planning, roadshow deck generation, visit memo, credit due diligence, financial statement analysis, industry analysis, equity penetration, credit application submission. -
AI Credit Underwriter – 7 skills: risk task planning, pre‑visit risk analysis, visit memo, admission‑rules scan, intake compliance check, related‑party detection, collateral risk management. -
AI Credit Risk Management Expert – 8 skills: policy environment analysis, industry risk rule generation, risk extraction, risk chain‑of‑thought generation, cross‑modal verification, loan monitoring, post‑loan management, large‑exposure concentration monitoring. -
AI Insurance Agent – 12 skills: customer profiling, gap analysis, client consulting, plan generator, objection handling, policy explanation, health disclosure guide, application check, renewal assistance, daily review, product wiki, social media marketing. -
AI Claims Specialist – 8 skills: claim registration, document processing, coverage analysis, medical review, claim adjustment, fraud detection. -
AI Underwriter (Insurance) – 5 skills: health verification, medical assessment, recording inspection, conclusion interpretation, follow‑up outreach. -
AI Research Analyst – 34 skills: global finance brief, earnings commentary, announcement analysis, institutional research outline, company deep analysis, industry deep analysis, macro data brief, policy analysis, strategy daily, sector allocation, market sentiment, factor tracking, backtesting, fixed‑income overnight brief, local‑government debt analysis, convertible bond valuation, and more. -
AI Investment Advisor – 9 skills: stock fund flow, quote analysis, technical analysis, shareholder structure, multi‑factor stock screening, multi‑factor fund screening, comparable company analysis, fund diagnosis, A‑share hotspot discovery. -
AI Data Scientist – 10 skills: data profiling, univariate analysis, feature analysis, XGBoost modeling, hyperparameter tuning, model explanation, segment modeling, DeepModel ensemble, model comparison, and auto‑experiment.
Each skill is defined with: name, trigger conditions, functional description, input/output specifications, and boundary conditions (what it doesn’t do). They are designed to be composable – a single agent can chain multiple skills to handle complex, multi‑step workflows.
The Next Three Years: From Strategic Consensus to Scaled Deployment
Three Bold Assertions
1. AI will move from “assistant” to “core capability definer.” By 2029, agentic AI will be embedded in end‑to‑end credit origination, product design, and real‑time risk hedging. The gap between institutions will be defined by AI capability density.
2. Tokens will become the “digital electricity” of finance. Inference costs have dropped >99% in three years. Even after the 2026 price corrections, 1M tokens cost about 0.7. As Jevons’ paradox kicks in, usage will explode – China’s daily token calls already exceeded 140 trillion by March 2026, growing 40% in three months. The winners will be those who manage token efficiency as a strategic resource.
3. AI employees will become the primary value‑creation force. By 2029, the crossover point – agents will handle the bulk of standardized execution, data analysis, client interaction, and risk monitoring. Humans will upgrade to strategic decision‑makers, AI trainers, relationship owners, and ethics overseers. Productivity will no longer be measured by headcount, but by the scale, density, and collaboration efficiency of the AI workforce.
Four‑Stage Adoption Path
-
Stage 1 – Awareness & Preparation (0–6 months): build executive consensus, assess infrastructure gaps, form new teams (prompt engineers, AI architects, skill developers). -
Stage 2 – Pilot & Validation (6–12 months): choose 1–2 high‑value, low‑risk scenarios; develop 3–5 core skills; run grayscale tests; quantify ROI. -
Stage 3 – Expansion & Deepening (12–24 months): scale to 20+ scenarios, 50+ core skills, and multiple agent roles working together. -
Stage 4 – Scale & Evolution (24–36 months): cover 80%+ of core business scenarios, 200+ skills, and enable autonomous skill evolution.
Five Critical Success Factors
-
Executive strategic patience – this is a marathon; early ROI may be modest, but the compounding effect is massive. -
Business‑technology co‑ownership – domain experts must co‑design skills; tech teams must iterate at daily, not monthly, cadence. -
Dynamic security‑innovation balance – implement layered guardrails, keep human‑in‑the‑loop for sensitive actions, and gradually expand autonomy only after proven reliability. -
Built‑in evolution mechanisms – isolation sandboxes, grayscale rollouts, and full skill lifecycle management are not optional. -
Ecosystem mindset – open skill development interfaces, participate in industry standards (MCP, A2A), and partner with technology providers rather than trying to own everything.
Summary, One‑Page Overview, and FAQ
Key Takeaways
-
The agent era is already here – measured by revenue, product velocity, and capex. -
Three engines (models, agents, ecosystem) feed a positive feedback loop – early movers accumulate compounding advantages. -
Moving to “AI‑native” requires four mechanisms – LUI, Hooks, Skills, and CLI – not just API wrappers. -
A general agent + pluggable skills is the winning architecture – 80% cost savings, natural cross‑domain collaboration, and faster evolution. -
Ten essential elements define a financial‑grade agent – from resilient execution to controlled evolution. -
100+ skills across ten roles are already defined and open‑sourced – they are a starting point, not a final set. -
Three‑level self‑evolution (agent, skill, model) ensures the system gets smarter without losing control. -
The next three years will see AI become the core capability differentiator – tokens will be a strategic resource, and AI employees will drive the majority of value.
One‑Page Quick Reference
FAQ
Q1: What’s the real difference between a general agent and a specialized one?
Specialized agents replicate basic capabilities for each scenario – expensive and hard to coordinate. A general agent shares one base and loads different skills. For 20 scenarios, specialized costs 5× more. For cross‑domain tasks, specialized agents actually degrade performance by 39%–70%.
Q2: How is a financial skill different from a regular API call?
A skill is a complete package: trigger conditions, step‑by‑step instructions, tool bindings, output schema, and embedded compliance rules. It encodes expert judgment, not just a function call.
Q3: How do you prevent a skill from introducing compliance risks during evolution?
Every proposed skill change goes through an isolation sandbox, automated and human validation, and grayscale rollout. If any metric degrades, it’s rolled back instantly. All changes are auditable.
Q4: How can we manage token costs at scale?
Smart routing sends simple queries to small models, moderate ones to medium models, and only the hardest 5% to large models. Combined with caching, context pruning, and output length control, costs drop by 60%–70% while maintaining quality.
Q5: Does “controlled evolution” conflict with “autonomous learning”?
No. The agent learns continuously, but with layered controls – personality tweaks happen instantly, skill changes go through a validation pipeline, and model updates require extended testing. Every change is reversible.
Q6: Which role should we automate first?
Start with “high‑frequency, standardized, low‑risk” scenarios – e.g., portfolio health checks, market briefs, or customer service triage. Build 3–5 skills, validate ROI, then expand.
Q7: How do we connect legacy systems without rewriting them?
Use CLI and MCP bridges. A CLI bridge can turn a shell script into a callable tool with just a descriptor file. MCP provides protocol‑grade integration for APIs – no code changes needed in the legacy system.
Q8: Are the 100+ skills the final set?
No. They are the starting point. Financial products, regulations, and client needs keep changing – skills must keep up. The entire library is open‑sourced to encourage continuous contribution and refinement by the industry.
This article is based on the comprehensive industry playbook “Financial Industry Agent 100 Skills” published by Alibaba Cloud’s New Finance team. The full open‑source skill library is available at github.com/allyun/qwen-dianjin/DianJin-SKILLS.
